Repository navigation
gh-85267: Improvements to inspect.signature __text_signature__ handling - #98796
Conversation
…handling This makes a couple related changes to inspect.signature's behaviour when parsing a signature from `__text_signature__`. First, `inspect.signature` is documented as only raising ValueError or TypeError. However, in some cases, we could raise RuntimeError. This PR changes that, thereby fixing python#83685. (Note that the new ValueErrors in RewriteSymbolics are caught and then reraised with a message) Second, `inspect.signature` could randomly drop parameters that it didn't understand (corresponding to `return None` in the `p` function). This is the core issue in python#85267. I think this is very surprising behaviour and it seems better to fail outright. Third, adding this new failure broke a couple tests. To fix them (and to e.g. allow `inspect.signature(select.epoll.register)` as in python#85267), I add constant folding of a couple binary operations to RewriteSymbolics. (There's some discussion of making signature expression evaluation arbitrary powerful in python#68155. I think that's out of scope. The additional constant folding here is pretty straightforward, useful, and not much of a slippery slope) Fourth, while python#85267 is incorrect about the cause of the issue, it turns out if you had consecutive newlines in __text_signature__, you'd get `tokenize.TokenError`. Finally, the `if name is invalid:` code path was dead, since `parse_name` never returned `invalid`.
|
The |
| except NameError: | ||
| raise RuntimeError() | ||
| raise ValueError | ||
|
|
There was a problem hiding this comment.
These now get caught on L2254 and acquire an error message there
JelleZijlstra
left a comment
There was a problem hiding this comment.
Looks good, just a few comments.
Can we backport this to the bugfix branches? All of the changes here seem like they fix buggy behavior, so I'm leaning to yes.
Co-authored-by: Jelle Zijlstra <[email protected]>
Co-authored-by: Jelle Zijlstra <[email protected]>
|
Thanks! Made the requested changes. I agree that this seems backport-able |
✅ Deploy Preview for python-cpython-preview canceled.
|
commented
Dec 12, 2022
|
#21104 did a subset of what this PR is changing, fixed the merge conflict. |
commented
Dec 21, 2022
|
Thanks @hauntsaninja for the PR, and @JelleZijlstra for merging it 🌮🎉.. I'm working now to backport this PR to: 3.10, 3.11. |
commented
Dec 21, 2022
|
Sorry, @hauntsaninja and @JelleZijlstra, I could not cleanly backport this to |
commented
Dec 21, 2022
|
Sorry @hauntsaninja and @JelleZijlstra, I had trouble checking out the |
commented
Dec 21, 2022
|
@hauntsaninja do you want to do the manual backports? |
commented
Dec 21, 2022
|
GH-100392 is a backport of this pull request to the 3.11 branch. |
commented
Dec 21, 2022
|
GH-100393 is a backport of this pull request to the 3.10 branch. |
commented
Dec 21, 2022
commented
Dec 21, 2022
|
Thank you! |
This makes a couple related changes to inspect.signature's behaviour when parsing a signature from
__text_signature__.First,
inspect.signatureis documented as only raising ValueError or TypeError. However, in some cases, we could raise RuntimeError. This PR changes that, thereby fixing #83685.(Note that the new ValueErrors in RewriteSymbolics are caught and then reraised with a message)
Second,
inspect.signaturecould randomly drop parameters that it didn't understand (corresponding toreturn Nonein thepfunction). This is the core issue in #85267. I think this is very surprising behaviour and it seems better to fail outright.Third, adding this new failure broke a couple tests. To fix them (and to e.g. allow
inspect.signature(select.epoll.register)as in #85267), I add constant folding of a couple binary operations to RewriteSymbolics.(There's some discussion of making signature expression evaluation arbitrary powerful in #68155. I think that's out of scope. The additional constant folding here is pretty straightforward, useful, and not much of a slippery slope)
Fourth, while #85267 is incorrect about the cause of the issue, it turns out if you had consecutive newlines in
__text_signature__, you'd gettokenize.TokenError.Finally, the
if name is invalid:code path was dead, sinceparse_namenever returnedinvalid.